Same Layout, Different Count: What Number Words Do Past Ten in Text-to-Image Models
Organizations: University of Colorado Boulder, Boulder, Colorado, USA
Abstract
Ask SD3.5-Medium for twelve of an object and then, from the same starting noise, for twenty-five. The picture keeps its layout, but the count changes about as much as it would under a new random seed, and in no consistent direction. We study what number words do once a request passes about ten objects, in five open text-to-image models. Holding the object and the starting noise fixed, we change one thing at a time, and people who never see the request count 3,088 of the images. Every model follows the number for small crowds, but only up to a point, its reach: SD3.5-M and FLUX.1-dev stop following it soon after ten, while FLUX.2 [klein] and Qwen-Image follow it loosely up to twenty-five. Beyond SD3.5-M's reach, the coarse part of the starting noise fixes the composition of the picture but not how many objects it holds: redrawing only the noise finer than an eighth of the image keeps the layout as well as a new number does, and already moves the count two-thirds as far. The number word still reaches the denoiser and acts in its first steps, but nothing we tried gives it a clear direction. Shown instead as a noised grid in the starting latent, the count is drawn: people find it in 59% of images, against 4% when it is only told. Past ten, an exact-match score barely separates a model that follows the number loosely from one that ignores it, and a relative-error score can even prefer a constant guess of sixteen. A same-noise test with a few requests per object tells them apart.
Figures & tables
| measure | what it is, and how it reads |
| order | share of differing same-noise pairs in which the larger request gives more objects: 1 always, .5 no direction; a loose estimator .72–.80 at 12–25 |
| change | mean between two images from the same noise; a picture counted twice by people: .12 |
| layout kept | correlation of thumbnails; past SD3.5-M’s reach, a new seed .10, a new number .70 |
| effective scatter | the loose estimator’s that orders a range as the model does; the reach: at most 0.45 |
| same-noise order | effective | exact | |||
| model (text encoder) | 2–8 | 12–25 | with 30 | scatter a | 12–30 |
| FLUX.1-dev (T5, CLIP) | .90 | .47 [.30, .64] | .17 | 3 [0.74, 3] | 0% |
| SD3.5-M (CLIP, T5) | .80 | .57 [.43, .69] | .49 | 1.55 [0.53, 3] | 4% |
| SDXL (CLIP) | .89 | .65 [.53, .78] | .60 | 0.69 [0.32, 3] | 3% |
| FLUX.2 [klein] (Qwen3) | .89 | .73 [.62, .84] | .44 | 0.42 [0.24, 0.89] | 13% |
| Qwen-Image (Qwen2.5-VL) | .99 | .78 [.65, .90] | .59 | 0.33 [0.17, 0.68] | 21% |
| layout | SAM 3, 64 types | people, 16 types | |||
| change from the same noise | kept | 2, 4, 8 | 12, 20, 30 | 4, 8 | 12, 20, 30 |
| a new number | .70 | .63 | .51 | .37 | .43 |
| a new seed | .10 | .29 | .49 | .25 | .53 |
| the number in digits | .76 | .25 | .48 | .52 | |
| “photograph” for “photo” | .18 | .42 | |||
| redraw above 2 cycles | .40 | .25 | .49 | ||
| SD3.5-M, same noise (SAM 3 order) | 2–8 | 12–30 |
| plain prompt (two seeds) | .86 | .50 [.45, .55] |
| word amplified, | .88 | .41 [.28, .54] (.15) |
| number in CLIP only / in T5 only | .85 / .85 | .47 / .55 |
| reward-tuned: Flow-GRPO / DiffusionNFT a | .95 / .92 | .38 / .32 |
| count LoRA / shuffled-count LoRA | .92 / .91 | .58 / .44 |
| count LoRA, people’s counts | .98 | .43 [.26, .61] |
| For a few dozen object types, draw four or five widely spaced requests past ten from one seed per type. | |
| Count the images with people, or with a counter checked against people. | |
| Report the same-noise order next to a loose estimator read through the same counting noise. | |
| From the same noise, add a new seed and an unrelated edit: a number that steers moves the crowd further than they do, and toward the request. |
| test | model | types | seeds | what changes, from the same noise | start |
| word ( Sec. C ) | SD3.5-M / | S1 | 3 / 1 | the request, 2–30 | 1 |
| SD3.5-M / | U2 | 2 | the request, 2–30 | 1 | |
| SDXL, FLUX.1-dev | S2 | 1 | the request, 2–30 | 1 | |
| FLUX.2 [klein] | U2 | 1 | the request, 2–30 | 1 | |
| Qwen-Image | 16 of U2 | 1 | the request, 2–30 | 1 | |
| asking differently ( Sec. C.3 ) | SD3.5-M | U2 | 1–4 | digits, layout sentence, guidance, best of four | 1 |
| study | requests | pictures | exact | within 10% | Spearman | ICC |
| first | all | 960 | .55 | .61 | .59 | .52 [.45, .59] |
| 2–8 | 256 | .72 | .73 | .74 | .60 [.43, .75] | |
| 12–30 | 704 | .49 | .56 | .54 | .47 [.39, .55] | |
| second | all | 464 | .61 | .68 | .77 | .73 [.65, .81] |
| 2–8 | 64 | .86 | .86 | .72 | .67 [.32, .92] | |
| 12–30 | 400 | .57 | .65 | .76 | .72 [.64, .80] |
| R1 = R2 | SAM 3 = people | ||||
| stratum | pictures | all | 12–30 | exact | bias |
| SD3.5-M plain | 144 | .67 | .55 | .67 | .8 |
| SDXL plain | 144 | .32 | .13 | .31 | 1.7 |
| FLUX.1-dev plain | 128 | .66 | .53 | .55 | 2.0 |
| SD3.5-M plain ∗ | 144 | .55 | .44 | .54 | .9 |
| self-planting ∗ | 144 | .63 | .63 | .51 | 2.7 |
| loose estimator, counted like the model | |||||
| model | measure | people | , matched | , matched | , one single count |
| FLUX.1-dev | order 12–30 | .36 [.23, .49] | .74 [.65, .82] ∗ | .81 [.74, .88] ∗ | .70 [.60, .78] ∗ |
| slope 12–30 | .36 [ .75, .02] | .91 [.52, 1.27] ∗ | .91 [.60, 1.19] ∗ | .83 [.22, 1.38] ∗ | |
| lift 12–30 | 1.2 [ 2.5, 0.3] | 1.8 [ 2.8, 7.2] | 4.0 [ 1.2, 10.3] | 1.6 [ 2.8, 6.6] | |
| lift 20–30 | 0.0 [0.0, 0.0] | 1.2 [ 3.6, 7.3] | 2.7 [ 3.1, 10.4] | 1.0 [ 3.1, 7.3] | |
| SD3.5-M | order 12–30 | .54 [.41, .65] | .74 [.65, .82] ∗ | .81 [.73, .87] ∗ | .70 [.60, .79] ∗ |
| slope | order | ||||
| Model | counter | 2–8 | 12–30 | 2–8 | 12–30 |
| SD3.5-M (S1, three seeds) | SAM 3 | .53 [.46, .59] | .19 [ .40, .03] | .88 | .49 [.42, .55] |
| Grounding DINO | .43 [.36, .49] | .68 [ .84, .53] | .84 | .31 [.27, .35] | |
| OWLv2 | .44 [.37, .50] | .17 [ .38, .02] | .82 | .48 [.42, .54] | |
| SD3.5-M (S1) | SAM 3 | .61 [.52, .69] | .38 [ .75, .02] | .90 | .53 [.44, .61] |
| Grounding DINO | .59 [.52, .66] | .67 [ .96, .38] | .90 | .39 [.31, .48] | |
| requested | 2 | 3 | 4 | 6 | 7 | 8 | 10 | 11 | 12 | 13 | 14 | 16 | 17 | 20 | 23 | 25 | 29 | 30 |
| SD3.5-M | .81 | .67 | .51 | .13 | .09 | .09 | .02 | .02 | .07 | .01 | .01 | .00 | .00 | .00 | ||||
| SD3.5-M | .79 | .60 | .62 | .38 | .21 | .21 | .08 | .15 | .10 | .04 | .06 | .00 | .02 | .00 | ||||
| SDXL | .52 | .27 | .21 | .15 | .06 | .06 | .02 | .08 | .06 | .00 | .00 | .02 | .02 | .00 | ||||
| FLUX.1-dev | .69 | .54 | .31 | .10 | .13 | .02 | .02 | .00 | ||||||||||
| FLUX.2 [klein], U2 | .66 | .52 | .20 | .09 | .11 | .08 | .00 | .05 | .02 | |||||||||
| SD3.5-M plain, U1 | .09 | .01 | .01 | .01 | .01 | .01 | .00 |
| order | slope | exact | ||
| SD3.5-M | people 2–8 | .95 [.89, 1.00] | .57 [.47, .69] | .33 [.25, .39] |
| people 12–30 | .49 [.34, .64] | .12 [ .55, .28] | .04 [.00, .10] | |
| SAM 3 2–8 | .88 [.75, .97] | .46 [.23, .63] | .33 [.25, .39] | |
| SAM 3 12–30 | .64 [.47, .81] | .10 [ .28, .47] | .00 [.00, .00] | |
| SD3.5-M | people 2–8 | .80 [.66, .93] | .37 [.16, .53] | .36 [.28, .45] |
| people 12–30 | .54 [.41, .65] | .01 [ .51, .47] | .04 [.00, .08] |
| , people | , SAM 3 | AIC AIC | people’s medians at 12, 16, 20, 25, 30 | |
| U2, 16 types (fifth count) | ||||
| FLUX.1-dev | 4.2 [2.8, 6.8] | 2.8 [2.0, 4.4] | 10.5 | 4, 8, 3.5, 2, 1.5 |
| SD3.5-M | 6.5 [3.0, 11.2] | 4.3 [2.5, 9.5] | 3.7 | 6.5, 6.5, 6, 8.5, 4.5 |
| SDXL | 10.9 [7.4, 17.8] | 5.1 [2.5, 10.8] | 3.8 | 14, 10, 14, 18, 18 |
| Qwen-Image | 13.5 [10.0, 15.1] | 10.3 [2.3, 17.2] | 20.7 | 12, 12, 15.5, 16, 15 |
| FLUX.2 [klein] | 15.2 [10.3, 20.7] | 5.4 [2.7, 19.4] | 6.4 | 12, 15, 19, 20, 9 |
| order | slope | exact | lift | none | |||
| model | counter | 2–8 | 12–30 | 12–30 | 12–30 | 12–30 | 12–30 |
| FLUX.2 [klein] | SAM 3 | .81 | .67 [.60, .74] | .29 [.05, .54] | .050 | .030 | .39 |
| Grounding DINO | .89 | .45 [.39, .51] | .55 | .028 | .009 | ||
| OWLv2 | .79 | .51 [.44, .58] | .13 | .025 | .009 | ||
| consensus | .87 | .59 [.44, .74] | .37 | .025 | .012 | ||
| people (16 types) | .89 | .62 [.53, .72] | .12 | .125 | |||
| order | slope | exact | within 10% | |||
| model | counter | 2–8 | 12–30 | 12–30 | 12–30 | 12–30 |
| Qwen-Image | SAM 3 | .70 | .78 [.67, .88] | .49 | .138 | .20 |
| Grounding DINO | .92 | .47 [.29, .64] | .44 | .138 | ||
| OWLv2 | .71 | .56 [.39, .73] | .01 | .088 | ||
| consensus | .78 | .72 [.51, .90] | .12 | .125 | ||
| people | .99 | .70 [.57, .82] | .23 | .213 | .28 | |
| area | box % | order | slope | exact | median | ||
| prompt | ratio | 12 / 30 | 2–8 | 12–30 | 12–30 | 12–30 (%) | 12 / 20 / 30 |
| SAM 3 | |||||||
| plain | 1 | 6.8 / 10.8 | .89 | .46 [.38, .53] | .20 | 2.5 | 8 / 6 / 4.5 |
| “seen from far away” | .47 | 4.2 / 3.3 | .88 | .54 [.46, .61] | .04 | 2.2 | 10 / 9 / 9 |
| “tiny” | .72 | 4.2 / 4.8 | .85 | .65 [.57, .72] | .27 | 2.5 | 7 / 7 / 7.5 |
| Grounding DINO | |||||||
| median box area (% of image), by request | slope of log area | |||||||||
| model | 2 | 4 | 8 | 12 | 16 | 20 | 25 | 30 | 2–8 | 12–30 |
| SD3.5-M, U2 (two seeds) | 27.2 | 16.2 | 10.7 | 7.2 | 8.0 | 7.7 | 6.0 | 9.2 | 0.70 [ 0.82, 0.58] | 0.18 [ 0.05, 0.41] |
| FLUX.2 [klein], U2 | 26.5 | 14.7 | 7.4 | 5.0 | 4.5 | 2.7 | 2.2 | 2.0 | 0.75 [ 0.95, 0.54] | 0.73 [ 1.15, 0.34] |
| SDXL, U2 | 15.9 | 10.7 | 4.6 | 3.8 | 3.1 | 3.1 | 2.5 | 2.0 | 0.75 [ 0.98, 0.52] | 0.55 [ 0.85, 0.26] |
| SDXL, screened set | 16.0 | 10.9 | 5.0 | 2.9 | 2.7 | 2.2 | 2.0 | 2.0 | 0.81 [ 1.00, 0.61] | 0.20 [ 0.65, 0.32] |
| Qwen-Image, U2 (16 types) | 29.4 | 15.3 | 9.1 | 6.4 | 6.7 | 5.4 | 6.4 | 4.5 | 1.03 [ 1.41, 0.77] | 0.12 [ 0.33, 0.09] |
| model, set | 2–8 | 12–30 | 12–25 | pairs with 30 | with 30 12–25 | 20/25 |
| FLUX.1-dev | .90 [.79, .98] | .36 [.23, .49] | .47 [.30, .64] | .17 [.03, .35] | .29 [ .51, .08] | .38 (3/5/8) |
| SD3.5-M | .80 [.66, .93] | .54 [.41, .65] | .57 [.43, .69] | .49 [.29, .67] | .08 [ .29, .14] | .67 (8/4/4) |
| SDXL | .89 [.76, .98] | .63 [.52, .74] | .65 [.53, .78] | .60 [.42, .77] | .05 [ .27, .14] | .57 (8/6/2) |
| Qwen-Image | .99 [.97, 1.00] | .70 [.57, .82] | .78 [.65, .90] | .59 [.40, .76] | .19 [ .37, .02] | .45 (5/6/5) |
| FLUX.2 [klein] | .89 [.80, .96] | .62 [.53, .72] | .73 [.62, .84] | .44 [.23, .65] | .30 [ .54, .04] | .50 (8/8/0) |
| SD3.5-M , S1 | .95 [.89, 1.00] | .49 [.34, .64] | .57 [.41, .72] | .36 [.16, .56] | .21 [ .38, .05] | .82 (9/2/5) |
| model, set | 2/4 | 4/6 | 6/8 | 8/12 | 12/16 | 16/20 | 20/25 | 25/30 |
| FLUX.1-dev | 1.00 (14/0/2) | .69 (9/4/3) | .92 (11/1/4) | .42 (5/7/4) | .67 (8/4/4) | .45 (5/6/5) | .38 (3/5/8) | .12 (1/7/8) |
| SD3.5-M | .87 (13/2/1) | .88 (7/1/8) | .56 (5/4/7) | .71 (10/4/2) | .58 (7/5/4) | .50 (7/7/2) | .67 (8/4/4) | .27 (3/8/5) |
| SDXL | .77 (10/3/3) | .92 (12/1/3) | 1.00 (8/0/8) | .71 (10/4/2) | .58 (7/5/4) | .71 (10/4/2) | .57 (8/6/2) | .50 (6/6/4) |
| Qwen-Image | .94 (15/1/0) | 1.00 (15/0/1) | 1.00 (14/0/2) | .87 (13/2/1) | .86 (6/1/9) | .77 (10/3/3) | .45 (5/6/5) | .36 (5/9/2) |
| FLUX.2 [klein] | .93 (13/1/2) | .93 (14/1/1) | .60 (6/4/6) | 1.00 (12/0/4) | .75 (9/3/4) | .75 (12/4/0) | .50 (8/8/0) | .33 (5/10/1) |
| SD3.5-M , S1 | 1.00 (12/0/4) | .75 (9/3/4) | 1.00 (12/0/4) | 1.00 (13/0/3) | .46 (6/7/3) | .54 (7/6/3) | .82 (9/2/5) | .17 (2/10/4) |
| effective scatter (annotators, matched) | reach | ||||||
| model, set | 2–8 | 12–25 | 12–30 | sliding, | sliding, | adjacent pairs | cap |
| FLUX.1-dev | 0.30 [ 0.05, 0.63] | 3 [0.74, 3] | 3 [ 3, 3] | 8 [8, 20] | 8 [4, 20] | 8 [4, 25] | 4.2 [2.8, 6.8] |
| SD3.5-M | 0.55 [0.20, 1.31] | 1.55 [0.53, 3] | 3 [0.78, 3] | 16 [6, 25] | 16 [4, 16] | 25 [6, 30] | 6.5 [3.0, 11.2] |
| SDXL | 0.25 [ 0.05, 0.64] | 0.69 [0.32, 3] | 0.90 [0.45, 3] | 12 [8, 30] | 12 [4, 25] | 30 [2, 30] | 10.9 [7.4, 17.8] |
| Qwen-Image | 0.05 [ 0.05, 0.18] | 0.33 [0.17, 0.68] | 0.55 [0.29, 1.71] | 25 [20, 30] | 25 [20, 25] | 25 [20, 30] | 13.5 [10.0, 15.1] |
| FLUX.2 [klein] | 0.36 [0.17, 0.60] | 0.42 [0.24, 0.89] | 0.99 [0.52, 3] | 25 [16, 25] | 25 [6, 25] | 25 [6, 30] | 15.2 [10.3, 20.7] |
| requests | counter | number / seed | order | signed shift | |
| 2, 4, 8 | SAM 3 | 2.01 [1.59, 2.63] | .89 [.83, .94] | .42 | |
| Grounding DINO, OWLv2, cons. | 2.04, 1.85, 2.94 | .82, .86, .89 | .34, .44, .38 | ||
| 2, 3, 4, 6, 8 | SAM 3 | 1.73 [1.42, 2.17] | .86 [.81, .91] | .31 | |
| 12, 20, 30 | SAM 3 | 0.96 [0.76, 1.19] | .46 [.37, .55] | .10 [ .22, .01] | |
| Grounding DINO, OWLv2, cons. | 1.00, 0.92, 0.85 | .36, .42, .29 | .26, .18, .21 | ||
| nine, 12–30 | SAM 3 | 0.89 [0.80, 0.99] | .50 [.45, .55] | .01 [ .07, .05] |
| requests | number | seed | number / seed | noise-free | order | |
| people | 2, 4, 8 | .52 | .19 | 2.77 [1.87, 4.09] | 2.04 [0.96, 4.10] | .92 [.80, 1.00] |
| 12, 20, 30 | .48 | .53 | 0.90 [0.60, 1.29] | 0.93 [0.41, 1.28] | .52 [.35, .69] | |
| SAM 3 | 2, 4, 8 | .57 | .35 | 1.63 [1.16, 2.59] | .82 [.67, .95] | |
| 12, 20, 30 | .50 | .69 | 0.72 [0.42, 1.06] | .51 [.34, .69] |
| SD3.5-M, U2 | value | spelling | seed | value order |
| SAM 3 | .51 [.39, .63] | .48 [.38, .62] | .49 [.40, .60] | .41 [.32, .51] |
| Grounding DINO | .53 | .44 | .54 | .40 |
| OWLv2 | .60 | .57 | .59 | .43 |
| consensus | .33 | .37 | .39 | .25 |
| thumbnails | box maps | |||
| pair | 12, 20, 30 | 2, 4, 8 | 12, 20, 30 | 2, 4, 8 |
| a new number | .71 [.68, .75] | .62 [.59, .66] | .58 [.52, .64] | .55 [.49, .60] |
| digits | .76 [.73, .79] | .64 [.58, .70] | ||
| a new seed | .10 [.04, .17] | .19 [.12, .26] | .32 [.26, .39] | .48 [.43, .53] |
| another type | .02 [ .02, .07] | .06 [.00, .12] | .20 [.15, .25] | .30 [.25, .35] |
| number seed | .61 [.55, .68] | .43 [.36, .51] | .24 [.18, .30] | .08 [.03, .12] |
| mean | |||||||||
| model, requests | counter | number | seed | digits | edit | number / seed | number / edit | edit / seed | order of number |
| SD3.5-M, 2, 4, 8 | SAM 3 | 0.63 | 0.29 | 0.25 | 0.18 | 2.17 [1.72, 2.81] | 3.50 [2.69, 4.84] | 0.62 [0.44, 0.82] | .91 [.85, .96] |
| Grounding DINO | 0.53 | 0.25 | 0.24 | 0.20 | 2.08 | 2.66 | 0.78 | .80 | |
| OWLv2 | 0.64 | 0.34 | 0.29 | 0.18 | 1.85 | 3.52 | 0.53 | .87 | |
| consensus | 0.58 | 0.18 | 0.15 | 0.10 | 3.25 | 5.53 | 0.59 | .89 | |
| SD3.5-M, 12, 20, 30 | SAM 3 | 0.51 | 0.49 | 0.48 | 0.42 | 1.03 [0.79, 1.30] | 1.22 [0.97, 1.56] | 0.84 [0.63, 1.12] | .41 [.32, .51] |
| mean | |||||||||
| model, requests | counter | number | seed | digits | edit | number / seed | number / edit | edit / seed | order of number |
| FLUX.2 [klein], 4, 8, 12 | SAM 3 | 0.57 | 0.40 | 0.38 | 0.23 | 1.41 [0.88, 2.28] | 2.43 [1.57, 4.52] | 0.58 [0.30, 0.96] | .89 [.78, .97] |
| Grounding DINO | 0.56 | 0.39 | 0.30 | 0.30 | 1.45 | 1.90 | 0.76 | .75 | |
| OWLv2 | 0.55 | 0.51 | 0.49 | 0.28 | 1.08 | 1.93 | 0.56 | .83 | |
| consensus | 0.50 | 0.24 | 0.23 | 0.13 | 2.11 | 3.90 | 0.54 | .93 | |
| FLUX.2 [klein], 20, 25, 30 | SAM 3 | 0.52 | 0.70 | 0.74 | 0.50 | 0.74 [0.45, 1.17] | 1.04 [0.61, 1.67] | 0.72 [0.51, 0.94] | .46 [.36, .55] |
| early steps ( ), 128 trajectories | estimate |
| value / spelling: twelve twenty-five over twelve “12” | 1.34 [1.25, 1.44] |
| value / spelling: two four over two “2” | 2.79 [2.54, 3.03] |
| value / spelling: twenty thirty over twenty “20” | 0.78 [0.72, 0.84] |
| size relative to guidance: 12 25, 12 “12” | 0.40, 0.33 |
| size relative to guidance: 2 4, 2 “2”; another object | 0.60, 0.25; 1.45 |
| 2 vs. 4, SAM 3 | 12 vs. 25, SAM 3 | 2 vs. 4 | 12 vs. 25 | ||||||
| steps | shift | toward ( ) | changed | shift | toward ( ) | changed | shift: Grounding DINO/ OWLv2/ cons. | shift: Grounding DINO/ OWLv2/ cons. | |
| 0–4 | 1.000 | .15 [.03, .27] | .76 (37) | .58 | .09 [ .30, .12] | .46 (41) | .64 | .14 / .10 / .11 | .17 / .11 / .30 |
| 5–9 | .953 | .16 [.03, .29] | .85 (26) | .41 | .03 [ .12, .19] | .68 (31) | .48 | .09 / .06 / .07 | .09 / .01 / .02 |
| 10–14 | .897 | .00 [ .09, .08] | .50 (14) | .22 | .08 [ .03, .23] | .64 (28) | .44 | .03 / .07 / .01 | .11 / .03 / .01 |
| 15–19 | .828 | .02 [ .09, .05] | .42 (12) | .19 | .05 [ .05, .18] | .59 (22) | .34 | .01 / .03 / .00 | .04 / .01 / .01 |
| 20–24 | .741 | .02 [ .05, .01] | .25 (4) | .06 | .04 [ .04, .17] | .40 (20) | .31 | .00 / .01 / .01 | .04 / .01 / .00 |
| people | SAM 3, same pictures | toward | |||||
| pair | steps | shift | toward ( ) | changed | shift | toward ( ) | diff. |
| 2 vs. 4 | 0–4 | .13 [ .09, .31] | .77 (13) | .81 | .39 [.13, .68] | .85 (13) | .08 |
| 10–14 | .02 [ .04, .07] | .67 (3) | .19 | .11 [ .01, .26] | .75 (4) | .08 | |
| 20–24 | .03 [.00, .08] | 1.00 (1) | .06 | .05 [ .16, .05] | .33 (3) | .67 | |
| 12 vs. 25 | 0–4 | .03 [ .43, .33] | .46 (13) | .81 | .18 [ .60, .18] | .44 (9) | .02 |
| 10–14 | .20 [ .06, .58] | .60 (10) | .63 | .03 [ .04, .11] | .80 (5) | .20 | |
| SAM 3 | |||||
| window (steps) | shift | toward (types) | changed | shift: Grounding DINO/ OWLv2/ consensus | |
| SDXL, two vs. four | |||||
| 2 vs. 4: every step | .41 [.27, .55] | .85 (54) | .84 | .38 / .43 / .35 | |
| swap 0–4 | .17 [.02, .31] | .74 (39) | .61 | .19 / .19 / .23 | |
| swap 5–9 | .16 [.06, .27] | .77 (31) | .48 | .09 / .14 / .18 | acts |
| swap 10–14 | .09 [.00, .17] | .76 (17) | .27 | .02 / .03 / .06 | acts |
| SAM 3 | |||||
| window (steps) | shift | toward (types) | changed | shift: Grounding DINO/ OWLv2/ consensus | |
| FLUX.2 [klein], two vs. four | |||||
| 2 vs. 4: every step | .26 [.10, .41] | .82 (50) | .78 | .43 / .27 / .25 | |
| swap 0–5 | .17 [.03, .31] | .83 (30) | .47 | .21 / .21 / .17 | acts |
| swap 6–11 | .16 [.10, .23] | .96 (24) | .38 | .16 / .11 / .12 | acts |
| swap 12–17 | .01 [ .07, .04] | .63 (8) | .13 | .03 / .01 / .01 | |
| request, push | counter | L-order | shift at | medians at |
| twelve, up | SAM 3 | .36 [.27, .46] | 0.23 [ 0.44, 0.03] | 8, 8, 8, 6, 4 |
| Grounding DINO, OWLv2, cons. | .50, .35, .32 | 0.09, 0.21, 0.12 | ||
| twelve, down | SAM 3 | .73 [.66, .79] | 0.59 [ 0.78, 0.40] | 8, 9, 8, 5, 2 |
| Grounding DINO, OWLv2, cons. | .66, .66, .77 | 0.36, 0.59, 0.64 | ||
| twenty-five, up | SAM 3 | .43 [.35, .52] | 0.10 [ 0.34, 0.14] | 6, 7, 6, 6, 4 |
| twenty-five, down | SAM 3 | .68 [.60, .75] | 0.60 [ 0.84, 0.35] | 6, 7, 3, 2, 2 |
| two vs four | twelve vs twenty-five | ||||||
| counter | CLIP | T5 | both | CLIP | T5 | both | |
| SD3.5-M | SAM 3 | .85 [.79, .92] | .85 [.78, .91] | .91 [.84, .97] | .47 [.38, .56] | .55 [.47, .64] | .53 [.41, .65] |
| Grounding DINO | .88 | .80 | .88 | .30 [.22, .38] | .42 | .31 | |
| OWLv2 | .80 | .74 | .89 | .45 | .55 | .46 | |
| consensus | .91 | .90 | .94 | .26 [.13, .42] | .61 | .34 | |
| FLUX.1-dev | SAM 3 | .63 [.38, .83] | .84 [.73, .93] | .83 [.69, .94] | .64 [.48, .78] | .36 [.15, .56] | .53 [.33, .71] |
| exact at 12, 20, 30 | SAM 3 | Grounding DINO | OWLv2 | consensus |
| said, 58 types (all 64) | ||||
| number word | .035 [.011, .063] (.031) | .011 (.016) | .029 (.026) | .023 (.021) |
| digits | .029 [.006, .055] (.026) | .006 (.005) | .006 (.005) | .006 (.005) |
| rows-and-columns sentence | .069 [.035, .105] (.062) | .069 (.062) | .035 (.031) | .046 (.042) |
| native resolution | .046 [.017, .083] (.042) | .035 (.031) | .052 (.047) | .029 (.026) |
| shown, 58 types, start at | ||||
| people, 12–30 | exact | within one | within 10% | recognisable | median |
| SD3.5-M , word (S1) | .04 [.00, .10] | .05 | .06 | .03 | .49 |
| SD3.5-M , word | .04 [.00, .08] | .09 | .10 | .04 | .33 |
| SD3.5-M, amplified word ( ) | .04 [.00, .08] | .08 | .08 | .03 | .46 |
| FLUX.1-dev, word (S2) | .08 [.02, .16] | .14 | .14 | .06 | .49 |
| SDXL, word (S2) | .04 [.00, .08] | .06 | .10 | .00 | 1.32 |
| SDXL, Gu et al . ’s first pass | .00 | .04 | .04 | .00 | .33 |
| SAM 3 | people, 16 types | ||
| exact at 12, 16, 20, 25, 30 (U2) | 64 types | exact | recognisable |
| SD3.5-M, same cells and noise | |||
| plain prompt | .022 | .04 [.00, .08] | .04 [.00, .08] |
| self-planting, grid | .470 | .59 [.40, .76] | .53 [.35, .70] |
| self-planting, scatter | .312 | ||
| given the grid’s boxes, from pure noise | |||
| 16 types | count only | yes or partly | recognisable |
| plain prompt ∗ | .04 [.00, .08] | .04 [.00, .08] | .04 [.00, .08] |
| self-planting | .50 [.33, .67] | .50 [.33, .67] | .44 [.28, .60] |
| GLIGEN ∗ | .46 [.31, .60] | .45 [.30, .60] | .18 [.09, .28] |
| InstanceAssemble | .65 [.56, .74] | .64 [.54, .73] | .38 [.26, .50] |
| self-planting, complete counts | .59 [.40, .76] | .59 [.40, .76] | .53 [.35, .70] |
| self-planting, partial counts | .39 [.19, .61] | .39 [.19, .61] | .33 [.16, .52] |
| start shows | object drawn | ||
| own cut-out | .531 [.43, .63] | .393 | .81 |
| another type’s cut-out | .345 [.26, .44] | .234 | .63 |
| mean-colour ellipses | .245 [.16, .33] | .276 | .54 |
| own another | .186 [.08, .29] | .159 | |
| own ellipses | .286 [.19, .38] | .117 | |
| another ellipses | .100 [ .01, .21] | .041 |
| planted kept | word | shift | ||||
| Model | set | conflict | agree | wins | to word | |
| SD3.5-M | S1 | .909 | .52 | .52 | .021 | .03 |
| SD3.5-M | S2 | .909 | .65 | .66 | .010 | .02 |
| SDXL | S2 | .881 | .59 | .57 | .007 | .08 |
| FLUX.1-dev | S2 | .933 | .58 | .65 | .035 | .02 |
| U1 | SAM 3 | consensus | vs. grid |
| plain prompt | .019 | .009 | .359 |
| best of two plain samples | .036 | – | .343 |
| self-planted start, same cut-out and object size | |||
| random placement | .199 | – | .180 |
| hole-free scatter | .273 | – | .105 |
| grid, random scale and offset | .343 | – | .036 |
| 64 types, 12–30 | GLIGEN | GLIGEN | InstanceAssemble | self-planting |
| SAM 3 | .158 | .152 | .324 | .406 |
| Grounding DINO | .188 | .212 | .457 | .351 |
| OWLv2 | .122 | .118 | .139 | .273 |
| consensus | .134 | .135 | .300 | .360 |
| complete counts (12, 16, 20, 25, 30) | .253 | .234 | .372 | .470 |
| partial counts (14, 19, 22, 27) | .039 | .049 | .264 | .326 |
| counter | planted | plain | planted plain | order, planted | none, planted |
| SAM 3 | .748 [.65, .84] | .055 | .69 [.59, .79] | .92 [.87, .96] | .23 |
| Grounding DINO | .672 [.58, .76] | .031 | .64 [.55, .73] | .80 [.72, .88] | .17 |
| OWLv2 | .434 [.33, .53] | .028 | .41 [.31, .51] | .84 [.77, .90] | .31 |
| consensus | .638 [.53, .74] | .028 | .61 [.50, .71] | .93 [.87, .97] | .21 |
| people (16 types) | .93 [.79, 1.00] | .13 | .80 [.66, .90] |
| exact, 4 and 8 | exact, 12–30 | output first | order, 12–30 | |
| counter | first / output | first / output | at 12–30 | first / output |
| SAM 3 | .164 / .422 | .009 / .075 | .07 [.03, .10] | .63 / .58 |
| Grounding DINO | .156 / .422 | .003 / .059 | .06 [.03, .08] | .48 / .47 |
| OWLv2 | .125 / .258 | .016 / .037 | .02 [ .01, .05] | .58 / .55 |
| consensus | .109 / .383 | .000 / .056 | .06 [.03, .08] | .53 / .54 |
| SAM 3, 12–30 | exact | within one | ratio |
| grids, planted at or given as boxes | |||
| self-planting, U1 / U2 / U3 | .378 / .406 / .366 | .443 / .464 / .433 | .85 / .88 / .85 |
| grid, random scale and offset (U1) | .343 | .429 | .80 |
| content: own / another / ellipses (U2) | .531 / .345 / .245 | .583 / .407 / .297 | .91 / .85 / .83 |
| GLIGEN / (U2) | .158 / .152 | .178 / .166 | .89 / .92 |
| InstanceAssemble (U2) | .324 | .343 | .94 |
T2I Generation
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.
See all 329 papers